Manav Garg

Raised data integrity 60% by taking a loosely defined brief with no specification through to production: rebuilt the ETL pipelines and data lake from scratch on S3, redesigned the ingestion schemas and enforced validation gates at every stage. Cut infrastructure cost 22% by profiling workload behaviour and scheduling, and sizing compute to actual usage rather than scaling it; removed 95% of production query errors by redesigning data layout and indexes around real access patterns. Lifted distributed model serving reliability 40% while cutting compute overhead 30%, migrating inference off a managed service onto a self hosted Kubernetes cluster that held under load the managed setup could not.

Backend AI Engineer at Vault22 (2026-03 – 2026-06)

• Raised data integrity 60% by taking a loosely defined brief with no specification through to production: rebuilt the ETL pipelines and data lake from scratch on S3, redesigned the ingestion schemas and enforced validation gates at every stage.
• Cut infrastructure cost 22% by profiling workload behaviour and scheduling, and sizing compute to actual usage rather than scaling it; removed 95% of production query errors by redesigning data layout and indexes around real access patterns.
• Lifted distributed model serving reliability 40% while cutting compute overhead 30%, migrating inference off a managed service onto a self hosted Kubernetes cluster that held under load the managed setup could not.
• Cut report turnaround 75% by embedding with an overseas customer team across three time zones, working out how their operations actually ran before writing anything, then shipping the reporting layer they needed.

Associate Data Scientist at IHX Pvt. Ltd. (A Perfios Company) (2025-01 – 2026-03)

• Owned the tariff digitization initiative end to end on healthcare claims data, taking a five day manual process down to a few hours by combining LLM extraction with document classification and validation gates at every stage.
• Reached 92% extraction accuracy on scanned and photographed clinical and billing documents by adapting SOTA models (YOLO, Donut), fine tuning and benchmarking them against each other, and running the annotation in Label Studio and error analysis that selected the production model.
• Segmented medical bills into recurring patterns using unsupervised KNN clustering, surfacing the structural groupings that drove how the pipeline handled each document type.
• Cut LLM hallucination over 70% with zero fine tuning by building RAG pipelines with structured prompting, output validation and guardrails, evaluated against real production failures rather than a benchmark.
• Sustained sub 200ms latency across 10,000+ daily requests by deploying, monitoring and maintaining production model serving end to end, including API design, queueing, logging and root cause debugging.
• Shipped a full digitization portal with cross functional teams across product, engineering and operations, owning the data cleaning and processing layer and explaining model limitations to non technical stakeholders.

AI/ML Intern at Intel Corporation (2024-05 – 2024-08)

• Achieved 94% accuracy classifying contract clauses and flagging anomalies with an end to end NLP pipeline on an event driven architecture over a high volume document stream.
• Reached 89% precision detecting fraud signals and subtle fine print deviations using semantic embeddings, scoped with the compliance team who acted on the output.

Bachelor of Technology in Computer Science Engineering – Vellore Institute of Technology (2021-08 – 2025-06)